Text Severity

The CQC require an automatic tool to analyse large numbers of online patient comments and detect significant negative changes in care quality. This poses challenges which distinguish it from standard text mining problems, including: 1. A highly unbalanced dataset: cases of interest will make up only a very small percentage of the data (c.0.5%); 2. A variable domain: the tool must cope with comments from a range of online sources, and be adaptable to social media in future; 3. A sensitive use case: the desired characteristics of the output (in particular, whether it is more important to avoid false positives or false negatives) will depend on how it is used.